Patients selected to participate in multimodal pain rehabilitation programmes in primary care−a multivariate cross-sectional study focusing on gender and sick leave

1 Introduction

Chronic pain is a common condition. A large study found that 19% of the population of Europe suffer from chronic pain of at least moderate severity [1]. Chronic pain frequently causes negative impacts on daily life. Worldwide, chronic pain in the neck and/or lower back are considered the leading cause of disability [2]. For the individual, chronic pain often leads to great suffering and is associated with impairment in physical and psychosocial functioning [3]. In Sweden, musculoskeletal pain is one of the most common reasons for sick leave [4], especially for women. A higher proportion of women than men are on sick leave [5]. Moreover, women are more frequently prescribed part-time sick leave than men [6], [7].

A multimodal/multidisciplinary rehabilitation programme (MMRP) is an evidence-based intervention for treatment of chronic pain conditions [8], [9], [10], [11] with mild to moderate effect depending on content. MMRP is generally performed as a group-based intervention with a bio-psycho-social approach. MMRPs include a combination of educational, physical and psychological interventions. A team of different health professionals, e.g. physician, physiotherapist, occupational therapist, social worker and psychologist, delivers a package of well-synchronized interventions over a period of several weeks, working together with the active patient to reach predetermined and important goals of the patient.

Previously, MMRPs have been delivered at specialist pain rehabilitation clinics in Sweden. In 2008, the Swedish government introduced a rehabilitation guarantee to enhance treatment, which included MMRPs for patients with chronic musculoskeletal pain delivered by primary care rehabilitation clinics, and to increase the rate of return to work as well as prevent sickness absence [12].

Fewer men than women participate in MMRPs [8]. This could be because of the higher prevalence of chronic pain in women than in men [13]. In addition, a contributing factor could be that women are more frequently recommended MMRP because of gender stereotypes [14], [15], [16]. An earlier study has shown that women and men to some extent display their problems in different ways when commencing MMRP in specialist care [17].

Patients with chronic pain do not constitute a homogenous group and there is consensus that chronic pain conditions are complex and require a bio-psycho-social approach of treatment and rehabilitation. Multivariate statistics can be useful for capturing and evaluating the complexity of the clinical presentation of chronic pain, and for identifying clinically meaningful subgroups of patients which facilitates the customising of MMRPs.

The aim of the study was to investigate the multivariate relationships between self-reported instruments in patients with chronic pain before taking part in an MMRP in primary care, and to have a special focus on gender and degree of sick leave.

2 Materials and methods

The design was a cross-sectional cohort study. The study was performed in two Swedish county councils, one in northern Sweden (Västerbotten) and one in southern Sweden (Östergötland) between August 2012 and September 2015. Patients from 12 MMRP teams in primary health care, in total 397 patients (339 women and 58 men) were consecutively recruited, and they filled in a questionnaire immediately prior to the start of MMRP. For background data, see Tables 1 and 2.

2.2 Statistics

Univariate and bi-variate statistics were performed using IBM SPSS version 22–23 (NY: IBM Corp. Released 2011). Multivariate statistics were performed using SIMCA-P 13.03 and SIMCA-P 15.02 (Umetrics Inc., Umeå Sweden).

Women and men, and patients with different degrees of sick leave (no sick leave, part-time sick leave and full-time sick leave) were compared using the χ² test, Mann-Whitney U-test, Kruskal-Wallis test, One-way ANOVA and Tukey’s Honest Significant Difference test. A 95% significance level was used.

Missing data were not replaced for the descriptive statistics. In the PCA and the OPLS analysis the estimation algorithm assigns zero weight to the missing observation. This corresponds to a missing completely at random (MCAR) assumption.

2.2.1 Principal component analysis

A principal component analysis (PCA) is generally used as a method to get a fast overview when dealing with a complex set of data. Hence, the aim of PCA is to reduce the dimensionality of the data and find the internal structure e.g. relations/correlations among a set of variables and transform the included variables into new uncorrelated components [34], [35]. These new components should explain the maximum possible quantity of variance in the data used. The components are extracted by the least squares criteria in order of decreasing importance. The strongest component is computed first and then a second one, which is orthogonal to the first and so on. Two to four components are generally enough to describe the relations in the data. In each component, it is possible to identify groups of observed variables with high correlations between variables in contrast to other variables. For variables included in the PCA, see Table 4. To enhance an easier interpretation of the results, all included variables were coded so that a higher score indicated larger problems (Table 4).

Table 4:

Variables included in the PCA analysis and loadings in the four components P1–P4; loadings ≤–0.25 and ≥0.25 are marked in bold type.

Variables (n=392)	P1	P2 (P2/P2SE*)	P3 (P3/P3SE*)	P4 (P4/P4SE*)
European quality of life – 5 Dimensions (EQ-5D) (a higher score indicates more problems) EQ-5D Index	0.28	0.14	−0.04	0.03
Pain-related disability (FRI)	0.28	0.22	−0.09	0.11
Hospital Anxiety and Depression Scale (HADS) (a higher score indicates more problems) Depression	0.28	−0.08	0.15	−0.16
European quality of life – 5 Dimensions (EQ-5D) (a higher score indicates more problems) EQ-VAS	0.27	0.07	−0.09	−0.11
Hospital Anxiety and Depression Scale (HADS) (a higher score indicates more problems) Anxiety	0.26	−0.22	0.16	−0.08
Chronic Pain Acceptance Questionnaire (CPAQ) (a higher score indicates less acceptance) Activity engagement	0.25	−0.06	−0.06	−0.11
Work Ability Score (higher score indicates less work ability) WAS	0.22	0.25	−0.16	−0.22
Pain Catastrophizing Scale (PCS) Rumination	0.20	−0.36	−0.02	0.18
Pain Catastrophizing Scale (PCS) Helplessness	0.24	−0.30	0.01	0.18
Pain Catastrophizing Scale (PCS) Magnification	0.17	−0.37	0.05	0.19
Chronic Pain Acceptance Questionnaire (CPAQ) (higher score indicates less acceptance) Pain willingness	0.17	−0.30	−0.05	0.17
Persistent pain	0.10	0.26	−0.15	0.16
Pain duration	0.00	0.21	0.32	0.08
Number of quadrants with pain	0.10	0.11	0.45	0.13
Number of earlier periods of pain	0.09	0.01	0.26	0.05
Pain sites vary	0.05	0.19	0.29	0.11
Number of pain sites	0.11	0.18	0.43	0.20
Degree of sick leave (higher score indicates greater degree of sick leave)	0.11	0.16	−0.14	−0.26
Life satisfaction Questionnaire (LiSat-11) (higher score indicates less satisfaction) Vocational dimension	0.16	0.01	0.07	−0.44
Life satisfaction Questionnaire (LiSat-11) (higher score indicates less satisfaction) Life dimension	0.23	−0.08	0.15	−0.33
Pain intensity (11-point numeric rating scale) last 7 days	0.20	0.17	−0.18	0.35
Pain intensity (11-point numeric rating scale) current	0.18	0.20	−0.22	0.35
Satisfaction with current level of functioning (higher score less satisfied)	0.22	0.01	−0.16	0.02
Chance to work within 6 month (higher score indicates less chance)	0.20	0.19	−0.06	−0.07
Worries about economy (higher score indicates more worries)	0.19	−0.01	0.06	−0.17
Number of doctor visits last year	0.11	0.01	−0.16	0.02
Expectations for rehabilitation (higher score indicates less expectations)	0.04	0.15	0.21	0.05
Godin-Shephard Leisure-Time Index (higher score indicates less activity)	0.09	−0.03	0.00	0.01
Frequency of sweat-inducing exercise (higher score indicates less activity)	0.09	0.01	−0.01	−0.03
R2	0.237	0.098	0.07	0.06
Q2	0.178	0.050	−0.005	0.01
Eigenvalue	6.87	2.83	1.97	1.87

1. R²=explanation of variance; Q²=prediction value.

2. Note that questionnaire scores are used in this analysis with a higher score indicating a negative state.

Each included variable was assigned a loading or component weight to each of the extracted components. Variables with loadings closer to 0 (irrespective sign) contributed less to the model. Variables with loadings ≤−0.25 and ≥0.25 were chosen (the interval is arbitrary) to represent the most important variables in the components, respectively. For each component, loadings with the same sign were positively correlated and variables with opposing signs were negatively correlated.

Factor rotation was not performed. Potential outliers were identified using score plots in combination with Hotelling’s T² and distance to model in X-space. No extreme outliers were identified.

R² labels the goodness of fit – the fraction of sum of squares of all the variables explained by a principal component [34]. Q² expresses the goodness of prediction – the fraction of the total variation of the variables that can be predicted by a principal component using cross validation methods [34]. R² and Q² are presented as adjusted.

To examine possible gender differences in the components, component scores were extracted and compared between groups using a 2-sample t-test.

2.2.2 Orthogonal partial least squares of latent structures (OPLS)

Partial least squares of latent structures (PLS) is a regression extension of PCA. PLS could be used as a generalized multiple regression method that can handle multiple collinear X and Y variables. OPLS is a modification of the traditional PLS and separates the systematic variations in X into two parts, one predictive correlated part and one uncorrelated part (orthogonal) [34].

Variables included in OPLS were the same as in the PCA (Table 4), with the addition of the variables age, gender, education, born in Sweden, Body Mass Index (BMI), permanent employee, jobseeker, number of doctor visits last year, living alone, living with wife/husband/partner, and living with children <18 years old.

To check the validity and the degree of overfit for the OPLS models, we used a permutation test [34], [36]. The test first estimated the OPLS model and its R2Y and Q2Y; then, with X fixed, the order of the elements in the Y-vector was randomly permuted 100 times. Each time a new OPLS model was fitted using X and the permuted Y, provided a reference distribution of R2Y and Q2Y for random data [36] (Fig. 1). This resulted in three permutation plots, each displaying the correlation coefficient between the original Y-variable and the permuted Y-variable on the X-axis versus the cumulative R² and Q² on the Y-axis, and a regression line. The intercept is a measure of the overfit [34].

Fig. 1:

Validate plot (after 100 permutations) of the Orthogonal partial least squares regression models (OPLS) for (a) Work ability score, (b) Degree of sick leave, and (c) Likelihood of working within 6 months. For validity, all prediction values (Q²) (squares) to the left are lower than the original points (square to the right), or the regression line of the Q² points (on the left) intersects the vertical axis at or below zero. R2Y=explanation of variance; Cum=cumulative.

3 Results

3.3 Regressions (OPLS)

To find important variables explaining the three work-related variables at baseline; WAS, degree of sick leave, and likelihood of working within 6 months, three OPLS regression analyses were performed. The validity and goodness of fit were significant according to the permutation test for the three OPLS models (Fig. 1) which means that the models were not over-fitted.

The OPLS of WAS explained 59% of the variation (R²Ycum). The prediction value was 52% (Q²) (Fig. 2). The most important variables associated with WAS were pain-related disability, degree of sick leave, quality of life, likelihood of working within 6 months, and acceptance (activity engagement) (Fig. 2).

Fig. 2:

Combined OPLS loading column plot with 95% confident interval bars. Relationship between the X-variable loadings (p) (light grey) and the Y-variable (Work ability score) loading (q) (dark grey) are displayed. Only significant variables in the model are shown. The height of the columns indicates the importance of each variable in the model, values closer to zero being less important irrespective of sign. Same sign at Y variable and X-variable means they are positively correlated. Y-variable with high loading means high correlation with the predictive component and X. Model variance R²Y=59% and prediction value Q²=52%. BMI=body mass index; CPAQ=chronic pain acceptance questionnaire; LW 6 months=likelihood of working within 6 months; EQ-5D=European Quality of life instrument; FRI=functional rating index; Godin index=Godin Shephard Leisure time Physical activity questionnaire; HADS=hospital anxiety and depression scale; LiSat=life satisfaction questionnaire; PCS=pain catastrophizing scale; WAS=work ability score.

The OPLS regression of degree of sick leave showed that work ability and pain-related disability were the most important regressors (Fig. 3). The full model explained 31% of the variance (R²Ycum) with a prediction value of 20% (Q²).

Fig. 3:

Combined OPLS loading column plot with 95% confident interval bars. Relationship between the X-variable loading (p) (light grey) and the Y-variable (Degree of sick leave) loading (q) (dark grey) is displayed. The height of the columns indicates the importance of each variable in the model, values closer to zero being less important irrespective of sign. Same sign at Y-variable and X-variable means that they are positively correlated. Y-variable with high loading means high correlation with the predictive component and X. Only significant variables are shown. Model variance R²Y =31% and prediction value Q²=20%. BMI=body mass index; CPAQ=chronic pain acceptance questionnaire; LW 6 months=likelihood of working within 6 months; EQ-5D=European Quality of life instrument; FRI=functional rating index; Godin index=Godin Shephard Leisure time Physical activity questionnaire; HADS=hospital anxiety and depression scale; LiSat=life satisfaction questionnaire; PCS=pain catastrophizing scale; WAS=work ability score.

The variables explaining most of the variations in likelihood of working within 6 months according to the OPLS regression analysis were pain-related disability, quality of life, depression, and work ability (Fig. 4). The full model explained 31% of the variance (R²Ycum) with a prediction value of 28% (Q²).

Fig. 4:

Combined OPLS loading column plot with 95% confident interval bars. Relationship between the X-variable loading (p) (light grey) and the Y-variable (Likelihood of working within 6 months) loading (q) (dark grey) is displayed. The height of the columns indicates the importance of each variable in the model, values closer to zero being less important irrespective of sign. Same sign at Y-variable and X-variable means they are positively correlated. Y-variable with high loading means high correlation with the predictive component and X. Only significant variables are shown Model variance R²Y=31% and prediction value Q²=28%. BMI=body mass index; CPAQ=chronic pain acceptance questionnaire; LW 6 months=likelihood of working within 6 months; EQ-5D=European Quality of life instrument; FRI=functional rating index; Godin index=Godin Shephard Leisure time Physical activity questionnaire; HADS=hospital anxiety and depression scale; LiSat=life satisfaction questionnaire; PCS=pain catastrophizing scale; WAS=work ability score.

4 Discussion

The main results from the present study from primary care were:

1. The PCA captured four components which together explained nearly half of the variation (47%) in the investigated data set. This indicates that the instruments included in the questionnaire did in fact capture much of the complexity and variation in patients with chronic pain participating in MMRP in primary care, although there is room for improvement.

2. No gender differences could be seen in components P1, P2 and P4, except in component P3, where pain variables had the highest loadings indicating that women reported more persistent and widespread pain.

3. Degree of sick leave was not well explained by the instruments in the questionnaire.

Since chronic pain patients in primary care are not a homogenous group [8], it is of interest to describe and identify subgroups of patients and to study patterns in order to better design their rehabilitation. MMRP is often performed in groups of patients but should also contain individualized elements including specific goals.

The largest variation in the data was captured in the first component P1 and clearly indicated that psychological distress (anxiety and depressive symptoms) together with less acceptance in activity engagement were associated with low quality of life (QoL) aspects. The variables with high absolute loadings on P1 contained a lot of information and showed considerable variations across patients. The associations between psychological and physical consequences as well as quality of life consequences in chronic pain confirm results from other studies [18], [37], [38]. Kroenke et al. found that nearly half of the patients with chronic pain in primary care suffered from anxiety which in turn impacted on several domains of health-related QoL such as mental and physical impairment [38]. The associations among the important variables in component P1 in our study might not be unique for chronic pain patients as Löwe et al. found that comorbidities of depression, anxiety and somatization caused more functional impairment than each individual variable did in a US multicentre study on general patients in primary care [39]. However, the importance of the variables in component P1 was also reported in a combined focus group and survey study on patients with chronic pain [18]. When asked, the patients listed enjoyment of life, emotional well-being, fatigue, weakness, and sleep-related problems as the most important aspects of their life being impacted by chronic pain [18]. One interesting observation is that pain duration did not contribute significantly to P1 at all (low loadings) i.e. pain duration was not associated with the interplay between psychological distress, less acceptance in activity engagement, and QoL aspects.

The second component P2 was characterized by catastrophizing and low willingness to accept pain. Ability to work did not seem to be affected. However, pain was less persistent. Patients who are characterized by this might be included in MMRP because of the catastrophizing rather than pain characteristics. Catastrophizing has been shown to predict disability and pain in some studies while in other studies no predictive effects of catastrophizing have been found [40].

The OPLS analysis showed that the instrument WAS were better explained by the variables included in the questionnaire than degree of sick leave. Since the rehabilitation guarantee was introduced in Sweden in 2008 with the aim to reduce sick leave, evaluation of MMRP now focuses more on sick leave and work ability than before. Recently, a study of MMRP in specialist care based on the SQRP and the Swedish Social Insurance Agency database showed a decrease in sick leave allowance for 2 years after MMR [41].

Using sick leave as an outcome measure for rehabilitation of chronic pain is a debatable matter. Studies indicate that professionals in primary care working with MMRP experienced difficulties in managing work ability and return to work [16], [42]. They perceived a conflict between what was seen as the goal of healthcare and the aim of the rehabilitation warranty and were more prone to focus on psychological wellbeing, participation in everyday life [42] and quality of life rather than on return to work. The professionals were of the opinion that the interventions in MMRP do not promote return to work in the first instance but could be a secondary consequence [16], [42]. This could be because MMRP in primary care is a relatively new intervention.

It can be asked which instrument is most appropriate to use when evaluating work ability in connection to MMRP. Earlier studies have shown divergent results regarding the effect of MMRP on sick leave [32], [43]. Some of the differences can be attributed to whether the data are self-reported or not. However, sick leave is a complex phenomenon and might depend on factors outside the patient’s control and factors not affected by MMRP. For example, degree of sick leave could be dependent on the design of the health insurance [33]. Other factors outside the patient’s control which can have an effect on sick leave are work-related factors, socioeconomic status [44] and poor treatment by healthcare professionals [45].

In comparison, WAS seems to be better explained by the variables in our analysis than degree of self-reported sick leave. Work ability is a balance between human resources and work (type and environment) and is influenced by context of society, family and close community [31]. A review [46] identified factors associated with poor work ability such as lack of leisure-time vigorous physical activity, poor musculoskeletal capacity, older age, obesity, high mental work demand, lack of autonomy, poor physical work environment, and high physical work load. At the time this study was conducted, no intervention was directed towards vocational rehabilitation during the MMRP. The baseline questionnaire could be developed further to serve as a tool for professionals to detect work-related factors impacting on work ability and sick leave. In a review, the authors [47] claimed that multimodal rehabilitation should be combined with vocational rehabilitation if the aim is to increase return to work. Directives are given in the Swedish national guidelines on what modalities should be included in MMRP for patients with chronic pain [48] there might be a need for better competence in vocational rehabilitation in primary care.

There were no multivariate differences between women and men in this study with the exception of component P3. This indicates that women had more widespread long-lasting pain in this study as has also been reported elsewhere. In a previous study performed at specialist care level, Rovner et al. found small but significant effect sizes indicating that when both women and men had the same pain severity, women described higher activity level, pain acceptance and social support while men reported higher kinesiophobia, mood disturbances and lower activity level before MMRP [17]. It is of importance to further study impact of gender in MMRP with larger samples of men included.

Multivariate methods offer a good method for gender analysis since they highlight how variables impact through intersections. If MMRP in primary care were to focus more on return to work, it might be recommendable to add instruments about work-related risk factors that have been shown to contribute to musculoskeletal pain in the same way in both women and men [49]. Another risk factor for developing pain is the gendered division of housework. Housework is an important source of stress that gives rise to consequences in the form of reported functional somatic symptoms such as pain. Increased housework during a 12-year period was associated with increased reported somatic symptoms in one study [50] for both women and men. In the questionnaire in this study, housework was not measured. This might be considered in the future when assessing patients before the start of MMRP to better adapt the rehabilitation based on individual needs.

A limitation in the study is that there were dropouts in some of the variables (Tables 1 and 2). This was partly due to that patients did not respond but also in some cases that the data were incorrectly filled in.

According to Swedish guidelines there is a medical indication for MMRP in primary care if the patient has chronic pain that significantly limits the patient’s daily life, and if the patient has the potential to improve despite the pain. The patient should also have tried unimodal treatment e.g. pharmacological treatment or physiotherapy without reaching any noticeable effects [48]. Although the inclusion criteria for participating in MMRP were based on the Indications for multimodal rehabilitation in chronic pain [48] we cannot rule out that there could also have been local inclusion criteria that have impacted the study results.

Multivariate methods with an inductive approach could be useful in studies on complex phenomena like chronic pain. This study has resulted in information about the four components that can be useful for designing programmes and evaluating MMRPs in the future. A limitation in this study could be that MMRP is a relatively new intervention in primary care and because of limited experience, patients unsuitable for MMRP may have been included.

5 Conclusion

The questionnaire filled out by the patients prior to participation in MMRP in primary care captured much of the complexity of chronic pain (conditions), but there is room for improvement, e.g. regarding explanation of work-related factors.

These findings indicate the importance of identifying subgroups of patients with chronic pain before MMRP in order to design their rehabilitation plans and customize interventions. In the multivariate analysis, gender did not appear to be an important factor for how most patients answered the questions.